Diagnose a slow or failed image restore with DavOS
Applies to: Medulla
Version: All versions (depending on tool availability)
Environment: On-Premise
Category: Imaging / DavOS / Clonezilla / NFS
Updated: November 19, 2026
1. Overview
This procedure helps diagnose a slow or failed image restoration performed using DavOS and Clonezilla in Medulla.
Two tools can be used to identify potential performance issues:
| Tool | Where to run it | What it checks |
|---|---|---|
davos-diag |
On the workstation running DavOS, during the restore | Hardware, network, NFS (identity, sessions, retransmissions, and latency), ongoing restore, disk, throughput, kernel messages, and DavOS/Clonezilla/partclone logs. |
davos-diag-serveur.sh |
On the Medulla server | NFS server, versions and threads, saturation, identities of connected workstations, network throughput, disk read speeds, and summary of workstation reports. |
Each check displays a verdict:
- [OK]: Check passed.
- [ATTENTION] (warning): Possible anomaly or issue requiring verification.
- [PROBLEME] (problem): An anomaly detected or a critical threshold reached.
The workstation report also indicates a likely bottleneck: network link, decompression, partclone, disk write, or network/NFS/server.
Note: Aside from generating reports, both tools perform read-only operations. They do not modify workstations, server settings, or images. They run with low CPU and disk priority to minimize their impact on the restore process.
2. Prerequisites
rootaccess to the Medulla server.- For manual diagnostics,
rootSSH access to workstations running DavOS. - A restoration in progress to obtain representative throughput measurements.
- Multiple workstations being restored simultaneously to identify any potential NFS concurrency issues.
It is recommended to run the measurements during the restoration of the main partition, when the partclone progress bar is visible.
3. Availability of diagnostic tools
Diagnostic tools are provided with the medulla-davos-client package. On versions that include them, no additional downloads are required.
| Tool | Location | Command |
|---|---|---|
| Workstation diagnostics | /usr/sbin/davos-diag, built into the DavOS image |
davos-diag |
| Server diagnostics | /usr/sbin/davos-diag-serveur.sh |
davos-diag-serveur.sh |
3.1. Check for the presence of tools
On the Medulla server:
ls -l /usr/sbin/davos-diag-serveur.sh
unsquashfs -l /var/lib/pulse2/imaging/davos/fs.squashfs | grep davos-diag
The first command checks for the presence of the server diagnostics. The second command checks that the workstation diagnostics are integrated into the DavOS image.
3.2. Previous versions of DavOS
If the tools are not present, it is recommended to update the medulla-davos-client package.
Otherwise, the scripts can be downloaded and installed manually:
wget https://dl.medulla-tech.io/ma/davos-diag.sh
wget https://dl.medulla-tech.io/ma/davos-diag-serveur.sh
cp davos-diag.sh /var/lib/pulse2/imaging/logs/
cp davos-diag-serveur.sh /usr/sbin/
chmod +x /usr/sbin/davos-diag-serveur.sh
The workstation script is then accessible from DavOS at:
/mnt/logs/davos-diag.sh
In this case, replace the davos-diag command with:
bash /mnt/logs/davos-diag.sh
4. Manual diagnosis
Step 1 – Start the restore
From Medulla, start restoring an image to one or more workstations.
Wait for Clonezilla to begin copying the main partition; the partclone progress bar will be visible.
Step 2 – Run the diagnostic on a workstation
Connect via SSH to the workstation running DavOS with root privileges.
To perform a 60-second measurement, with a sample every 5 seconds:
davos-diag
To extend the measurement to 120 seconds:
davos-diag -d 120
To change the duration and interval:
davos-diag -d 120 -i 10
When multiple workstations are restoring simultaneously, run the diagnostic on two or three of them to compare their results.
4.1. Available options
| Option | Description |
|---|---|
-d SECONDES |
Measurement duration, 60 seconds by default. |
-i SECONDES |
Interval between measurements, default is 5 seconds. |
-w |
Wait for partclone to start—for up to 15 minutes—before taking a measurement. |
-L |
Collect logs only, without performing any measurements. Used particularly at the end of a restore. |
Step 3 – Run the server diagnostic
While the workstations are restoring, run the following on the Medulla server:
davos-diag-serveur.sh -n 6
In this example, six workstations are currently being restored.
The -n parameter allows you to specify the expected number of workstations and compare it with the observed NFS clients.
Step 4 – Retrieve the results
The reports are saved on the server in:
/var/lib/pulse2/imaging/logs/davos-diag/
A subfolder is created for each workstation and each run, as well as a folder for server diagnostics.
The main files are:
| File | Content |
|---|---|
rapport.txt |
Summary, checks, and verdicts. |
echantillons.csv |
Performance metrics collected periodically. |
brut/ |
Detailed command outputs and log copies. |
5. Automatic diagnostics
DavOS can automatically run davos-diag during each restore, without requiring an SSH connection to the workstations.
This feature is disabled by default and can be enabled from the Medulla server.
5.1. Enable automatic diagnostics
On the Medulla server:
touch /var/lib/pulse2/imaging/logs/davos-diag.enable
This command enables a 60-second measurement.
To request a 120-second measurement:
echo '-d 120' > /var/lib/pulse2/imaging/logs/davos-diag.enable
Only the -d and -i options, followed by a numeric value, are interpreted in this file.
5.2. Enable diagnostics on a specific workstation
It is also possible to enable diagnostics using the kernel parameter:
davos_diag=1
This parameter enables diagnostics for the specified workstation.
5.3. How automatic diagnostics works
When automatic diagnostics are enabled:
- DavOS boots normally.
- The Clonezilla restore process begins.
- The diagnostics wait for partclone to start.
- Performance is measured for the configured duration.
- The results are saved on the server.
- Logs are collected at the end of the restore.
Two folders are typically created for each workstation:
<workstation>-<date>
<workstation>-<date>-fin
The first contains the metrics collected during the restore. The second contains the logs collected at the end of the operation.
Note: The diagnostic runs in the background with low priority. The steps are time-limited, and diagnostic errors do not block the restore. At the end of the restore, DavOS waits a maximum of 20 seconds for the logs to be collected.
5.4. Disable automatic diagnostics
rm /var/lib/pulse2/imaging/logs/davos-diag.enable
Without the activation file or the kernel parameter davos_diag=1, DavOS operates normally without running the automatic diagnostics.
5.5. Previous versions of DavOS
If the DavOS image does not yet include the automatic diagnostics mechanism, update the medulla-davos-client package.
Otherwise, download the integration installer:
wget https://dl.medulla-tech.io/ma/medulla-davos-diag-install.sh
bash medulla-davos-diag-install.sh
The script performs a backup of the existing DavOS image, adds the diagnostics, and then rebuilds the image.
The backup is named according to the following format:
fs.squashfs.avant-diag-<date>
To revert to the previous image:
bash medulla-davos-diag-install.sh --rollback
Important: Manual integration modifies the DavOS image. It is recommended to perform this operation during a maintenance window and to test the boot process on a pilot workstation.
6. Read and interpret the reports
Start with the Synthèse (summary) section at the end of the report, which lists the [ATTENTION] and [PROBLEME] checks.
Next, check the Goulot (bottleneck) line in the “Analyse des mesures” (measurement analysis) block.
6.1. Checks performed on the workstation
| Block | Check | Anomalies looked for |
|---|---|---|
| Hardware | Memory, processor, BIOS. | Available memory less than 512 MB. |
| Network | Speed, duplex, errors, latency. | Link speed less than 1,000 Mbit/s, half-duplex, network errors, ping greater than 5 ms. |
| NFS | Version, NFSv4 identity, rsize. | NFSv4 identity not defined, rsize less than 256 KB. |
| Restoration | ocs-sr and partclone processes. | No restores in progress. |
| Disk | SMART, temperature, kernel errors. | Critical NVMe warning, temperature of at least 70 °C, disk errors. |
| Analysis | Sessions, retransmissions, NFS latency. | Renegotiations, timeouts, TCP reconnections, latency greater than 50 ms. |
| Kernel messages | Network interface card, firmware, storage, NFS. | Freezes, connection losses, thermal throttling. |
| Logs | DavOS, Clonezilla, partclone. | Mount failures, disk too small, checksum errors. |
6.2. Identifying the bottleneck
The script determines the bottleneck based on the averages of the measurements, in the following order:
| Verdict | Condition | Interpretation |
|---|---|---|
| Lien utilisé à X % (link used at X%) | Received throughput is greater than or equal to 80% of the link’s capacity. | The workstation is heavily utilizing its network capacity. |
| Goulot : décompression (bottleneck: decompression) | pigz or zstd is using at least 90% of a core. | Decompression may limit throughput. |
| Goulot : partclone (bottleneck: partclone) | partclone uses at least 90% of a core. | Partclone processing may limit restoration. |
| Goulot : écriture disque (bottleneck: disk write) | iowait is 20% or higher. | The target storage is the first thing to investigate. |
| Goulot : réseau / NFS / serveur (bottleneck: network / NFS / server) | None of the above. | Examine NFS communications and the server. |
Interpretation: The verdicts help guide the diagnosis. They do not always provide definitive proof of the cause of the slowdown. It is recommended to compare the measurements from the workstation and the server.
6.3. Checks performed on the server
The server report specifically indicates:
- Generic or duplicate NFSv4 identities.
- A number of connected computers lower than the number specified.
- Fewer than 16
nfsdthreads. - More than 10% of NFS requests waiting for a thread.
- Server disk wait time greater than or equal to 20%.
- An average throughput per workstation of less than 300 Mbit/s.
7. Frequently asked questions
7.1. A single workstation is fast, but multiple workstations are very slow. Why?
One possible cause is an NFSv4 identity or session issue.
In some versions of DavOS, NFS shares may be mounted before the workstation’s final name is assigned.
Several computers may then have a generic identity:
debian
localhost.localdomain
The diagnostic message may read:
- “Identité NFSv4 non définie” (NFSv4 identity not defined).
- “Sessions NFSv4 renégociées” (NFSv4 sessions renegotiated).
- “Nom générique” (generic name) on the server side.
The NFSv4 identity fix assigns each workstation an identity based on its MAC address:
davos-<MAC address>
Important: An identical hostname and an identical NFSv4 identity are two different concepts. A generic or duplicate identity should be investigated, but does not by itself indicate a session collision or a loss of throughput. Also check for renegotiations, retransmissions, and measured performance.
7.2. The report states “Lien négocié à 100 Mbit/s” (link negotiated at 100 Mbit/s)
The workstation is not negotiating a Gigabit connection.
Check the cable, the jack, the patch cord, the switch port, and the network driver.
ethtool <interface> | grep -E 'Speed|Duplex'
A 100 Mbit/s link limits the theoretical throughput to approximately 12.5 MB/s, before protocol overhead.
7.3. The report indicates “Goulot : décompression” (bottleneck: decompression)
The processor is devoting a significant portion of its resources to decompressing the image.
Gzip images using the -z1p option may be limited by the workstation’s decompression performance.
One optimization approach is to use zstd with the Clonezilla -z9p option, depending on the entity’s settings.
An existing image can be converted using:
ocs-cvtimg-comp
It is recommended to compare performance on a pilot workstation before rolling out the change.
7.4. The report indicates “Goulot : réseau / NFS / serveur” (bottleneck: network / NFS / server)
The workstation appears to be waiting for the data required for the restore.
Check the following in this order:
- NFSv4 identity and sessions.
- NFS retransmissions and timeouts.
- NFS latency.
nfsdthreads.- Server disk wait time.
- Network throughput from the server.
7.5. The report indicates NFS retransmissions or TCP reconnections
These events may result from network losses, overloaded equipment, a firewall interrupting connections, or a server that is no longer responding quickly enough.
Check the network counters:
ethtool -S <interface> | grep -iE 'err|drop|miss'
Check for NFS communication losses:
dmesg | grep -i 'not responding'
Compare the counters before and during the restore to distinguish between old and new errors.
7.6. The report indicates “Goulot : écriture disque” (bottleneck: disk write)
The target disk may be limiting the restore.
Possible causes include:
- A mechanical disk.
- A full SSD.
- An SSD experiencing thermal throttling.
- A failing drive.
See:
brut/nvme-smart.txt
brut/dmesg.txt
7.7. Should you use NFSv3 or NFSv4?
It is recommended to stick with NFSv4 when clients have a properly initialized identity.
NFSv4 primarily uses TCP port 2049 and simplifies network configuration.
NFSv3 does not use the NFSv4 client identity mechanism but requires additional RPC services, notably rpcbind and mountd, depending on the configuration.
In NFSv4.2, certain operations—notably READ_PLUS—may exhibit performance differences depending on the server. To isolate these, compare the results with a mount using vers=4.1.
7.8. The partclone throughput does not match the network throughput. Is this normal?
Yes. Partclone displays the throughput of decompressed data written to disk, whereas the network transmits compressed data.
For example, with an image compressed by a factor of about two:
| Measurement | Approximate value |
|---|---|
| Partclone throughput | 12 GB/min |
| Restored data throughput | 200 MB/s |
| Corresponding network throughput | 800 Mbit/s |
7.9. The report indicates “Aucune restauration en cours” (no restore in progress)
The diagnostic was run before or after the data was copied.
The hardware, NFS, and log checks are still useful, but the throughput measurements are not representative.
Rerun davos-diag while the main partition is being restored.
7.10. The server shows fewer connected computers than the number of workstations launched
There are several possible explanations:
- Some workstations have not yet mounted the NFS share.
- Some workstations have completed their restoration.
- Some clients have abnormal NFSv4 identities.
- Some connections are not visible at the time of diagnosis.
Examine the NFSv4 identities, sessions, and reports for the affected workstations.
8. Log location
On the workstation, logs are stored in memory and are cleared upon reboot. The diagnostic tool copies them to the brut/logs/ folder.
| Log | On the workstation | On the server | Content |
|---|---|---|---|
| DavOS | /var/log/davos.log |
davos-diag/<workstation>-<date>/brut/logs/ |
Boot, NFS mounts, NFS identity, and requested action. |
| Restoration | /var/log/davos_restorer.log |
debug_imaging/<workstation>/davos_restorer-<date>.log in case of an error |
ocs-sr command, Clonezilla output, and errors. |
| Capture | /var/log/davos_saver.log |
debug_imaging/<workstation>/ |
Image capture command, compression, and throughput. |
| Clonezilla | /var/log/clonezilla.log |
brut/logs/ |
Partition-by-partition restoration process. |
| Partclone | /var/log/clonezilla/partclone.log |
brut/logs/ |
Average throughput and read or checksum errors. |
| Kernel messages | dmesg |
brut/dmesg.txt |
Network, drivers, NVMe, and NFS errors. |
| Image | — | masters/<UUID>/Info-saved-by-cmd.txt and clonezilla-img |
Capture command and compression settings. |
Paths: Server paths are relative to /var/lib/pulse2/imaging/logs/, except for masters/, which is relative to /var/lib/pulse2/imaging/.
8.1. Quickly find restoration errors
grep -inE 'error|fail|traceback|abort|no space|cannot' /var/log/davos_restorer.log | tail
8.2. View partclone throughput
grep -oE 'Ave\. Rate: *[0-9.]+ *[GM]B/min' /var/log/clonezilla/partclone.log | tail -3
8.3. Examine kernel messages
dmesg | grep -iE 'link is up|nvme|nfs|firmware|hang' | tail -20
9. The workstation does not boot into DavOS
If the workstation fails to boot into DavOS, no workstation diagnostic report can be generated.
In this case, check:
- The messages displayed on the screen.
- The operation of the PXE boot process.
- The PXE and HTTP service logs.
- The download of the
vmlinuzkernel. - The download of
initrd.img. - The loading of
fs.squashfs.
10. Information to provide to Medulla support
To enable a thorough analysis, please provide:
- The complete diagnostic folder.
- The number of workstations restored simultaneously.
- The duration of the restoration.
- The name of the restored image.
To create an archive:
tar czf davos-diag-$(hostname)-$(date +%Y%m%d).tar.gz -C /var/lib/pulse2/imaging/logs davos-diag
The folder containing the reports is:
/var/lib/pulse2/imaging/logs/davos-diag/
Confidentiality: Reports may contain computer names, IP addresses, hardware information, and internal paths. Check their contents before sending.
11. Summary
The DavOS diagnostic helps identify the main causes of slow or failed restores:
- Insufficient network throughput.
- Abnormal NFSv4 identities or sessions.
- NFS retransmissions and latency.
- NFS server saturation.
- Decompression-related limitations.
- Partclone limitation.
- Insufficient target disk performance.
- Hardware or restoration errors.
The davos-diag and davos-diag-serveur.sh tools are included with the medulla-davos-client package in versions that support this feature.
Diagnostics can be run manually or automatically during restores.
Key point: To effectively diagnose a slow image restore, compare the reports from multiple workstations with that of the Medulla server. This approach helps distinguish workstation-specific limitations from network issues, NFS sessions, or shared storage problems.